Data Selectionedit
Selection of training or evaluation data under reliability constraints.
Data Selection is the process of choosing which examples are used for training, pruning, evaluation, or synthetic-data reuse. Selection can reduce cost and improve quality, but biased selection can also distort a model's view of the target distribution.
Research contextedit
Data selection is a shared problem in Data Centric ML, AI and networks, and Synthetic Data. In decentralized or siloed settings, selection is often local: each participant sees only part of the data and chooses examples according to local goals or constraints. That makes selection a networked problem rather than a purely statistical preprocessing step.
Connection to Qiao's workedit
Data selection appears in When Sample Selection Bias Precipitates Model Collapse, where biased local selection can worsen recursive synthetic-data training and make low-resource communities more vulnerable to tail-mode loss. In the unlearning papers, selection reappears as removal or reweighting: the model is changed by changing which data count.